GLM 5.3

Runs MCP tools

Z.ai’s flagship reasoning model, released August 2026 with open weights. It reuses GLM 5.2’s base model, so every gain comes from post-training, and it always thinks before it calls a tool.

Vendor
Z.ai
Released
18 August 2026
Context
1.31M tokens
Input
Text

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

MCP Evals

Can GLM 5.3 actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has GLM 5.3 drive each one, and reports where it got the wrong answer even though every call succeeded. Pick GLM 5.3 as the driver.

What GLM 5.3 is

GLM 5.3 is a 744B-parameter mixture-of-experts model with 40B active per token. Z.ai built it for complex software engineering and long-horizon agent work, and post-trained it heavily on security tasks. It takes text only; GLM 5.3 Flash is the multimodal tier. Z.ai launched it on its API and GLM Coding Plan on 14 August 2026 and published the weights on Hugging Face on 28 August under its own GLM-5.3 licence.

What Z.ai says

  • CyberGym: 84.5%, against 83.8% for Claude Mythos 5 and 83.6% for GPT-5.6 Sol.
  • DeepSWE v1.1: 66.9, against 72.7 for GPT-5.6 Sol.
  • Humanity’s Last Exam with tools: 62.5, against 64.5 for GPT-5.6 Sol.
  • GDPval-AA v2: 1769, against 1743 for Claude Fable 5.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • MCP Playground’s Agent Studio: paste your server URL above and run GLM 5.3 against it in the browser, with no Z.ai key
  • Claude Code on the GLM Coding Plan, which Z.ai offers as a way to drive GLM from Claude Code. Claude Code is an MCP client.
  • Z.ai’s OpenAI-compatible API as glm-5.3, so MCP clients that accept a custom OpenAI-compatible endpoint can point at it with a base-URL change

What to watch for

  • Thinking can no longer be disabled. Z.ai’s API rejects the disabled setting and takes low, high or max reasoning effort instead, defaulting to max. A client that turned thinking off for fast tool calls on GLM 5.2 will need changing.
  • At max effort each tool call waits on a full reasoning pass. If a multi-step chain feels slow, try high before you conclude the model is too slow for your server.
  • Z.ai’s security post-training is the headline, but the DeepSWE and Terminal-Bench numbers still trail GPT-5.6 Sol. For general agent work, compare it with GLM 5.3 Flash on your own tools.

Run GLM 5.2 and GLM 5.3 on the same four-step task against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. Same base model, different post-training, so any difference in the calls comes from the post-training alone. All the mock servers →

Available here

ModelModel ID
GLM 5.3z-ai/glm-5.3

Go deeper

Frequently asked questions

Is GLM 5.3 open source?
The weights are public on Hugging Face since 28 August 2026, in FP8 and BF16. The licence is Z.ai’s own GLM-5.3 licence, not MIT, so read it before shipping the model in a product.
Can I turn off thinking on GLM 5.3?
Not on Z.ai’s API. The disabled setting is no longer accepted. You choose a reasoning effort of low, high or max, and max is the default.
What is the difference between GLM 5.3 and GLM 5.3 Flash?
GLM 5.3 is the large text-only reasoning model. Flash is the smaller, faster tier and adds image and video input. Both carry a window of over 1M tokens, and both are available here.
Does GLM 5.3 support MCP?
Yes, through any MCP client that translates tools into function calls, including Claude Code on Z.ai’s GLM Coding Plan. You can run it against your own server in MCP Playground without a Z.ai key.

Sources

GLM 5.3 for MCP: Specs & Tool Calling — Test Free | MCP Playground